PerceptualHash is quite different on version 14.7.0 vs prior versions. Expected?

[ozziepeeps]

Hi, thanks again for a great library!

(PS this is a similar post to #1720 which you helped me with!)

I have observed that the PerceptualHash generated by version 14.7.0 is quite different to 14.6.0. I have created a repo that demonstrates what I'm seeing: https://github.com/ozziepeeps/MagickNET-perceptualhash-change

The repo includes a single (public domain) image asset. I have then written some unit tests and generated the perceptual hash on:

• Magick.NET-Q16-AnyCPU 14.0.0
• Magick.NET-Q16-AnyCPU 14.6.0
• Magick.NET-Q16-AnyCPU 14.7.0

All 3 package versions are present in the csproj for your convenience.

I observe that the PerceptualHash does not change at all between 14.0.0 and 14.6.0 and the unit test HashDifference_MagickNET_Q16_AnyCPU_14_0_0_VS_MagickNET_Q16_AnyCPU_14_6_0 demonstrates that the difference is 0.

However the unit test HashDifference_MagickNET_Q16_AnyCPU_14_6_0_VS_MagickNET_Q16_AnyCPU_14_7_0 demonstrates that the difference between 14.6.0 and 14.7.0 is much larger: 19.069.

I am not sure if this is intended. I've observed pretty stable / negligible changes across versions over my years of using the NuGet, but this one stood out so I wanted to highlight it just in case it wasn't intended.

Thanks!

[dlemstra]

Thanks for reaching out and asking this question. This is expected because there have been a lot of changes inside ImageMagick 7.1.2-0.. When ImageMagick/ImageMagick#3950 was build a lot of fixes where made in the underlying logic of this method.

[ozziepeeps]

Ok, thanks for the reply @dlemstra. I rely on the stability of the pHash for a bunch of business functions. Do you know if it's possible to have 14.7.0 compute a backward compatible hash? (I'm guessing not). Do you think it might be worthwhile me re-computing my library of hashes on the new version (i.e. do you see it remaining stable going forward given your knowledge of developments in ImageMagick itself)? Any advice would be most appreciated!

[dlemstra]

There is no way to get a backwards compatibility version of that method. As far as I quick understand some of the mathematical methods for image statistics that are used by PerceptualHash contained errors. These errors have been fixed and that resulted in the difference that you are seeing now. The changes inside ImageMagick took several weeks to complete and I do think it's stable now. So I do think that it would be wise to do a complete recalculation. The ImageMagick release that contains these changes was made on the 13th of July so there is some risk that someone with a lot of math expertise will tell us we are wrong. But on the other side will be the person that is responsible for those changes and they have a lot of math expertise.

p.s. There was a small fix in 14.8.0 for arm64 incase you are using that.

[ozziepeeps]

Makes sense, appreciate your insight, as always!